<p>In motorcycle motorsport, engineering decisions are constrained by mass, stiffness, fatigue life, and thermal margins. In a comparable way, deploying large language models (LLMs) for edge or trackside assistance is limited by memory capacity, bandwidth, latency, and power. This paper presents a systems-engineering framework that relates structural lightweighting and post-training quantization (PTQ) through second-order sensitivity. In mechanics, safe material removal is governed by the stiffness matrix <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\textbf{K}\)</EquationSource> </InlineEquation>; in neural networks, safe precision reduction is governed by curvature information encoded in the loss Hessian <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\textbf{H}\)</EquationSource> </InlineEquation>. To ground this analogy, we combine a structural case study based on topology optimization of a Formula Student upright with an edge-oriented evaluation of low-precision LLM deployment using established PTQ pipelines (GPTQ and AWQ). The structural example illustrates how sensitivity-guided material removal preserves load paths and mechanical integrity, while the LLM study examines how sensitivity-aware quantization can reduce computational footprint while maintaining operational usefulness in telemetry-related inference. To assess deployment feasibility, we define two practical metrics: a Digital Factor of Safety (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\textrm{FoS}_{\textrm{dig}}\)</EquationSource> </InlineEquation>), based on a perplexity-derived integrity threshold, and Intelligence-per-Watt (IPW), which captures energy-aware inference efficiency. For a 32B-class operating point, INT4 weight-only deployment reduces memory footprint from 61 GB to 18 GB, improves throughput from 26.0 to 69.9 tok/s, and lowers measured power from 295 W to 165 W while remaining within the proposed integrity envelope. The contribution is methodological rather than algorithmic: we do not introduce a new quantization method, but a reproducible framework for analyzing when low-precision LLM deployment becomes feasible under motorsport-inspired resource constraints. The results support the practical relevance of stiffness-informed digital lightweighting, while also highlighting that the evidence is limited to the structural and computational case studies considered here.</p>

Structural optimization principles for edge AI in motorsport telemetry

期刊论文
出版模式：

开放获取
发表日期：

2026年5月1日

查看全文

Rubén Juárez Cádiz,
Fernando Rodríguez-Sela

摘要

In motorcycle motorsport, engineering decisions are constrained by mass, stiffness, fatigue life, and thermal margins. In a comparable way, deploying large language models (LLMs) for edge or trackside assistance is limited by memory capacity, bandwidth, latency, and power. This paper presents a systems-engineering framework that relates structural lightweighting and post-training quantization (PTQ) through second-order sensitivity. In mechanics, safe material removal is governed by the stiffness matrix \(\textbf{K}\) ; in neural networks, safe precision reduction is governed by curvature information encoded in the loss Hessian \(\textbf{H}\) . To ground this analogy, we combine a structural case study based on topology optimization of a Formula Student upright with an edge-oriented evaluation of low-precision LLM deployment using established PTQ pipelines (GPTQ and AWQ). The structural example illustrates how sensitivity-guided material removal preserves load paths and mechanical integrity, while the LLM study examines how sensitivity-aware quantization can reduce computational footprint while maintaining operational usefulness in telemetry-related inference. To assess deployment feasibility, we define two practical metrics: a Digital Factor of Safety ( \(\textrm{FoS}_{\textrm{dig}}\) ), based on a perplexity-derived integrity threshold, and Intelligence-per-Watt (IPW), which captures energy-aware inference efficiency. For a 32B-class operating point, INT4 weight-only deployment reduces memory footprint from 61 GB to 18 GB, improves throughput from 26.0 to 69.9 tok/s, and lowers measured power from 295 W to 165 W while remaining within the proposed integrity envelope. The contribution is methodological rather than algorithmic: we do not introduce a new quantization method, but a reproducible framework for analyzing when low-precision LLM deployment becomes feasible under motorsport-inspired resource constraints. The results support the practical relevance of stiffness-informed digital lightweighting, while also highlighting that the evidence is limited to the structural and computational case studies considered here.